A REFLEXIVE HI SPACE WITH THE HEREDITARY INVARIANT SUBSPACE PROPERTY SPIROS A. ARGYROS, PAVLOS MOTAKIS Abstract. A reflexive hereditarily indecomposable Banach space Xjgp is presented, such that for every Y inflnite dimensional closed subspace of Xjsp and every bounded linear operator T : Y ^ Y, the operator T admits a non-trivial closed invariant subspace. Introduction The invariant subspace problem asks whether every bounded hnear oper- ator on an infinite dimensional separable Banach space admits a non-trivial closed invariant subspace. A classical result of M. Aronszajn and K.T. Smith [9] asserts that the problem has a positive answer for compact op- erators. This result was extended by V. Lomonosov [18] for operators on complex Banach spaces that commute with a non-trivial compact operator. Recently G. Sirotkin [27] has presented a version of Lomonosov's theorem for real spaces. It is also known that the problem, in its full generality, has a negative answer. Indeed P. Enflo [13] and subsequently C. J. Read [23], [23] have provided several examples of operators on non-reflexive Banach spaces that do not admit a non-trivial invariant subspace. Also recently a non-reflexive hereditarily indecomposable (HI) Banach space Xk with the "scalar plus compact" property has been constructed [7]. This is a £oo space with separable dual, resulting from a combination of HI techniques with the fundamental J. Bourgain and F. Delbaen construction [lOj- As consequence, the space Xk satisfies the Invariant Subspace Property (ISP). All the above results provide no information in either direction within the class of reflexive Banach spaces. The aim of the present work is to construct a reflexive Banach space Xjgp with the hereditary ISP. Namely, every infinite dimensional closed subspace of Xjgp satisfies ISP, a property which is unknown for the aforementioned 2010 Mathematics Subject Classification: Primary 46B03, 46B06, 46B25, 46B45, 47A15 Key words: Spreading models, Strictly singular operators. Invariant subspaces. Reflex- ive spaces, Hereditarily indecomposable spaces 1 2 S.A. ARGYROS, P. MOTAKIS space Xk- It is notable that no subspace of Xjgp has the "scalar plus com- pact" property. More precisely, the strictly singular operator^ on every subspace Y of Xjgp form a non separable ideal (in particular, the strictly singular non-compact are non-separable). The space Xjgp is a hereditarily indecomposable space and every operator T G >C(Xjgp) is of the form T = XI + S with S strictly singular. We recall that there are strictly singular operators in Banach spaces without non-trivial invariant subspaces [25]. On the other hand, there are spaces where the ideal of strictly singular operators does not coincide with the corresponding one of compact operators and every strictly singular operator admits a non-trivial invariant subspace. The most classical result in this direction, due to V. Milman [19j, concerns the strictly singular operators in -L^'[0, 1],1 ^ p < cxD,C[0, 1]. This is a consequence of Lomonosov-Sirotkin theorem and the fact that the composition TS is a compact operator, for any T, S strictly singular operators, on any of the above spaces. In [2J, Tsirelson like spaces satisfying similar properties are presented. The possibility of constructing a reflexive space with ISP without the "scalar plus compact" property emerged from an earlier version of [2]. The following describes the main properties of the space j£jgp . Theorem. There exists a reflexive space jtjgp with a Schauder basis {en}n<=N satisfying the following properties. (i) The space Xjgp is hereditarily indecomposable. (ii) Every seminormalized weakly null sequence {x„}„gN has a subse- quence generating either £i or cq as a spreading model. Moreover every infinite dimensional subspace Y of Xjgp admits both ii and cq as spreading models. (iii) For every Y infinite dimensional closed subspace of Xjgp and every T G >C(y, Xjgp), T = Xly^^^^ + S with S strictly singular. (iv) For every Y infinite dimensional subspace of Xjgp the ideal S{Y) of the strictly singular operators is non separable. (v) For every Y subspace of Xjgp and every Q, S, T in S(Y) the operator QST is compact. Hence for every T € S(Y) either = or T commutes with a non zero compact operator. (vi) For every Y infinite dimensional closed subspace of X and every T € >C(y), T admits a non-trivial closed invariant subspace. In par- ticular every T ^ XIy, for A G M admits a non-trivial hyperinvariant subspace. It is not clear to us if the number of operators in property (v) can be reduced. For defining the space Xjgp we use classical ingredients like the A bounded linear operator is called strictly singular, if its restriction on any infinite dimensional subspace is not an isomorphism. HEREDITARY INVARIANT SUBSPACE PROPERTY 3 coding function a, the interaction between conditional and unconditional structure, but also some new ones which we are about to describe. In all previous HI constructions, one had to use a mixed Tsirelson space as the unconditional frame on which the HI norm is built. Mixed Tsirelson spaces appeared with Th. Schlumprecht space [26], twenty years after Tsirelson construction p8]. They became an inevitable ingredient for any HI construction, starting with the W.T. Gowers and B. Maurey celebrated example [16j, and followed by myriads of others etc. The most signif- icant difference in the construction of Xjgp from the classical ones, is that it uses as an unconditional frame the Tsirelson space itself. As it is clear to the experts, HI constructions based on Tsirelson space, are not possible if we deal with a complete saturation of the norm. Thus the second ingredient involves saturation under constraints. This method was introduced by E. Odell and Th. Schlumprecht |20j.[21j for defining heterogeneous local structure in HI spaces, a method also used in [2]. By saturation under constraints we mean that the operations {^,Sn) (see Re- mark [T3]) are applied on very fast growing families of averages, which are either a-averages or /3-averages. The a-averages have been also used in [20], [21], while /3-averages are introduced to control the behaviour of special functionals. It is notable that although the a, /3-averages do not contribute to the norm of the vectors in Xjgp , they are able to neutralize the action of the operations {-^jSn) on certain sequences and thus cq spreading mod- els become abundant. This significant property yields the structure of X^^p described in the above theorem. Let us briefly describe some further structural properties of the space X •^isp • The first and most crucial one is that for a (n, e) special convex combina- tion (see Definition II. 9p Yli^zp CiXi, with {xi}i£F a finite normalized block sequence, we have that W^CiXiW ^ — + 12e This evaluation is due to the fact that the space is built on Tsirelson space and differs from the classical asymptotic ii HI spaces (i.e. [3], [8]) where seminormalized (n,e) s.c.c. appear everywhere. A consequence of the above, is that the frequency of the appearance of RIS sequences is significantly increased, which yields the following. Every strictly singular operator maps sequences generating cq spreading models to norm null ones. Furthermore for every two strictly singular operators T, S : Xjgp — > Xjgp such that TS is non-compact and every weakly null sequence such that {TSxn}n is not norm convergent, the following holds. Every spreading model generated by a subsequence of {TSxn}n is cq. Combining the above properties we conclude property (v) of the above theorem. 4 S.A. ARGYROS, P. MOTAKIS We thank G. Costakis for bringing to our attention G. Sirotkin's paper 1. The norming set of the space Xjgp In this section we define the norming set W of the space Xjgp . This set is defined with the use of the sequence {Sn}n which we remind below and also families of 5„-admissible functionals. As we have mentioned in the introduction, the set W will be a subset of the norming set Wt of the Tsirelson space. The Schreier families. The Schreier families is an increasing sequence of families of finite subsets of the naturals, first appeared in [ij, inductively defined in the following manner. Set So = {{n} : n G N} and 5i = {F C N : #F ^ minF}. Suppose that 5„ has been defined and set = {F C N : F = Uj^^Fj, where Fi < • • • < G 5„ and k ^ min Fi } If for n,m G N we set 5„ * 5m = {F C N : F = U^=iF,-, where Fi < • • • < Ffc G Sm and {miuFj : j = 1, . . . ,k} £ then it is well known that Sn * Sm = Sji+m ■ Notation. A sequence of vectors xi < ■ ■ ■ < in cqo is said to be 5,^- admissible if {minsuppxj :i = l,...,A;}G5„. Let G C Coo- A vector / G G is said to be an average of size s{f) = n, if there exist fi, . . . , fd G G,d ^ n, such that / = H \- fd)- A sequence {fj}j of averages in G is said to be very fast growing, if /i < /2 < • • •, s{fj) > (maxsupp/j_i)2 and s{fj) > for j > 1. The coding function. Choose L = {ik ■ k G N},£i > 2 an infinite subset of the naturals such that: (i) For any A: G N we have that ik+i > 2^^^ and (ii) Yjk=i ^ < Tooo- Decompose L into further infinite subsets Li,L2. Set Q = {((/i,ni),...,(/m,nm)) :mGN,{nfc}^^i CN,/i < ... </mGcoo with fk{i) G Q, for 2 G N, /c = 1, . . . , m} Choose a one to one function a : Q — >• L2, called the coding function, such that for any ((/i, ni), . . . , {fm,nm)) € Q, we have that '7((/i,ni), . . . , {fm,nm)) > 2""" • maxsupp/m Remark 1.1. For any n G N we have that #L n {n, . . . , 2^"} ^ 1. The norming set. The norming set W is defined to be the smallest subset of Coo satisfying the following properties: 1. The set {±e„}„gN is a subset of W, for any / G we have that — / G W, for any f £W and any / interval of the naturals we have that If€W and HEREDITARY INVARIANT SUBSPACE PROPERTY 5 W is closed under rational convex combinations. Any / = ie^ will be called a functional of type 0. 2. The set W contains any functional / which is of the form / = ^ Sj=i '^j ; where {aj}j^i is an 5„-admissible and very fast growing sequence of a- averages in W. If / is an interval of the naturals, then g = izlf is called a functional of type !„, of weight w{g) = n. 3. The set W contains any functional / which is of the form f = ^ X]j=i f^j > where {Pj}'^^^ is an 5.„-admissible and very fast growing sequence of /?- averages in VF. If / is an interval of the naturals, then g = ±If is called a functional of type I/3, of weight ■w{g) = n. 4. The set W contains any functional / which is of the form / = ^ Sj=i /? ' where {fj}j=i is an 5i-admissible special sequence of type Iq, functionals. This means that w{fi) G Li and w{fj) = a{{fi,w{fi)),. . . , {fj-i,w{fj-i))), for j > 1. If / is an interval of the naturals, then g = 'tif is called a func- tional of type II with weights (^{g) = {w{fj) : ran /j fl / 7^ 0}. We call an a-average any average a ^ W oi the form a = ^ Si=i /?' ^ ^ n, where /i < • • • < /d € W. We call a /3-average any average /3 € of the form /3 = ^ Si=i /? ' ^ ^' where /i, • • • , /d € are functionals of type II, with disjoint weights w{fj)- In general, we call a convex combination any f W that is not of type 0, Iq,, 1/3 or II. For X G Coo define ||x|| = sup{/(a;) : / G W} and Xjgp = (coo(N), || • \\). Evidently Xjgp has a bimonotone basis. One may also describe the norm on Xjgp with an implicit formula. Indeed, for some x G Xjgp , we have that ^ d 1 |jx|| = max{||x||o, ||x||//, sup{— ^ H-EjxH^J, sup{— ^ ||Sjx||f^.}} where the inner suprema are taken over all re G N, all ^^-admissible intervals {Ej}j^^ of the naturals and ki < ■ ■ ■ < such that kj > (max£'j_i)^ for j>l- By ||x||77 we denote ||x||// = sup{/(x) : / G is a functional of type II} whereas for j G N, by ||x||" we denote ||x||^ = sup{a(x) : a G is an a-average of size s{a) = j} Similarly, by ||x||^ we denote ||x||^ = sup{/3(x) : /? G is a /3-average of size s(/3) = j}. Remark 1.2. Very fast growing sequences of a-averages have been consid- ered by E. Odell and Th. Schlumprecht in [20], [21] and were also used in 6 S.A. ARGYROS, P. MOTAKIS [2j. However, /3-averages are a new ingredient, introduced to control the be- haviour of type II functionals on block sequences. The /^-averages can also be used to provide an alternative and simpler approach of the main result in [21]. As we have mentioned in the introduction, the ||2:||^, \\z\\j, which are averages, do not contribute to the norm of the vector z. On the other hand, the {II ■ lljlj, {II • have a significant role for the structure of the space •^isp • Remark 1.3. The norming set W can be inductively constructed to be the union of an increasing sequence of subsets {Wm}m=o °f '^oo, where Wq = {-^n}neN and if has been constructed, then set VF^+i to be the closure of Wm under taking a-averages, W^_^_l to be the closure of W^^^ under taking type Iq, functionals, W^_^_i to be the closure of W^_^_i under taking type 1/3 functionals, W^+i to be the closure of VF^+i under taking type II functionals, VF^+i *° closure of Wj^_|_]^ under taking /^-averages and finally Wm+i to be the closure of W^^j^^ under taking rational convex combinations. Tsirelson space. Tsirelson's initial definition [2^ of the first Banach space not containing any £p, 1 ^ p < oo or cq, concerned the dual of the so called Tsirelson norm which was introduced by T. Figiel and W. B. Johnson [15] and satisfies the following implicit formula. 1 \\x\\t = max{||x||o, sup{-^ ||£;jx||r}} j=i where x € cqo and the inner supremum is taken over all successive subsets of the naturals d ^ Ei < ■ ■ ■ < E^. Tsirelson space T is defined to be the completion of (cqo, || • ||t)- In tbe sequel by Tsirelson norm and Tsirelson space we will mean the norm and the corresponding space from |15j . As is well known, a norming set Wt of Tsirelson space is the smallest subset of Coo satisfying the following properties. 1. The set {±en}neN is a subset of Wt, for any / G Wt we have that — / € Wt-, for any / G Wt and any E subset of the naturals we have that Ef G Wt and Wt is closed under rational convex combinations. 2. The set Wt contains any functional / which is of the form f = \ X]j=i fji where {/j}^=i is a Si admissible sequence in Wt- Remark 1.4. The following are well known facts about Tsirelson space. (i) The norming set Wt can be inductively constructed to be the union of an increasing sequence of subsets {W^jJ^^g of cqo, in a similar manner as above. (ii) The set VF^, which is the smallest subset of cqo satisfying the follow- ing properties, also is a norming set for Tsirelson space. HEREDITARY INVARIANT SUBSPACE PROPERTY 7 1. The set {+e„}„gN is a subset of W^, for any / G we have that — / € VF-p and for any / S and any E subset of the naturals we have that Ef G W^. 2. The set contains any functional / which is of the form / = h X]j=i /j) where {fj}j^i is a <Si admissible sequence in W/p. Remark 1.5. It is easy to check that the norming set Wt of Tsirelson space is closed under {-^^Sn) operations, namely for any /i < • • • < in Wt iS„-admissible, the functional ^ Yl'j=i fj ^ ^t- This explains that the normine set W of the subset of Wt- Therefore Tsirelson space is the unconditional frame on which the norm of Xjgp is built. As we mentioned in the introduction, Xjgp is the first HI construction which uses Tsirelson space instead of a mixed Tsirelson one. As it is shown in (see also |12j). an equivalent norm on Tsirelson space is described by the following implicit formula. For x G cqo set ^ 2d |||.t||| =max{||x||o, sup{- ^ i=i where the inner supremum is taken over all successive subsets of the naturals d ^ El < ■ ■ ■ < E2d- Then, for any {ckYj^^i ^^^^ following holds. n n n (1) W'^CkCkllT ^ III X^Cfeefcl ^ 3|| y^Cfcefcllr k=l k=l k=l Remark 1.6. A norming set ^^(t,|||-|||) foi' (T, ||| • |||) is also defined in a similar manner as Wt- Special convex combinations. Next, we remind the notion of the {n,e) special convex combinations, (see [1],[6],[8]) which is one of the main tools, used in the sequel. Definition 1.7. Let x = Ylk&F ^k^-k be a vector in cqo- Then x is said to be a (n,e) basic special convex combination (or a (n,e) basic s.c.c.) if: (i) F G 5„, Cfe ^ 0, for A; G F and Y.k(^F ^fc = 1- (ii) For any G C F,G & <Sn-i, we have that J2k€G < ^- The proof of the next proposition can be found in Chapter 2, Propo- sition 2.3. Proposition 1.8. For any M infinite subset of the naturals, any n G N and e > 0, there exists F C M, {ck}keF, such that x = '^k&pCkek is a (n,e) basic s.c.c. Definition 1.9. Let xi < - - - < Xm be vectors in cqo and ip{k) = minsuppx^, for /c = 1, . . . , m. Then x = X^fcLi CfcX^ is said to be a (n,e) special convex combination (or (n,e) s.c.c), if X^^Li Cfce^(fc) is a (n,e) basic s.c.c. 8 S.A. ARGYROS, P. MOTAKIS 2. The basic inequality In this section we prove the basic inequaUty for block sequences in Xjgp , with the auxihary space actuahy being Tsirelson space. This will allow us to evaluate the norm of (n, e) special convex combinations and it is critical throughout the rest of the paper. Definition 2.1. Let f £ W he a functional of type Iq or I^, of weight w{f ) = n, f = ^ Si=i Then, by definition, there exist Fi < ■ ■ ■ < Fp successive intervals of the naturals such that: (i) ul,F,={l,...,d} (ii) {minsupp fj : j £ Fi} G 5„_i, for i = 1, . . . ,p (iii) {minsupp /minF, : i = 1, . . . ,p} S 5i Set gi = ZjeF, /?■' iov i = 1, . . . ,p. We call {gjf^i a Tsirelson analysis of /. Remark 2.2. If / € is a functional of type I^ or I^ and is a Tsirelson analysis of /, then fi G W, is 5i-admissible and / = k Yl^=i /«' although {fi}^^i may not be a very fast growing sequence of a- averages or /3-averages. Moreover, if w{f) > 1, then fi is of the same type as / and w{fi) = w{f) — 1 for i = 1, . . . ,p. The tree analysis of a functional f € W. A key ingredient for evaluating the norm of vectors in Xjgp is the analysis of the elements / of the norming set W. This is similar to the corresponding concept that has occurred in almost all previous HI and related constructions (i.e. [3], [4], [7], [8]). Next we briefly describe the tree analysis in our context. For any functional f G W we associate a family {/aIasA, where A is a finite tree which is inductively defined as follows. Set /0 = /, where denotes the root of the tree to be constructed. If / is of type 0, then the tree analysis of / is {fz}. Otherwise, suppose that the nodes of the tree and the corresponding functionals have been chosen up to a height p and let A be a node of height |A| = p. If /a is of type 0, then don't extend any further and A is a maximal node of the tree. If fx is of type Iq, or Ip, set the immediate successors of A to be the elements of the Tsirelson analysis of fx. If fx is of type II, / = | Yl'j=i fji the immediate successors of A to be the {f,}%v If fx is a convex combination, which includes a-averages and /3-averages, fx = X^j=i Cj/j, set the immediate successors of A to be the {fj}'j=i- By Remark 11.31 it follows that the inductive construction ends in finitely many steps and that the tree A is finite. Remark 2.3. Let f £ W and {/aIasA be a tree analysis of /. Then for any A G A not a maximal node, such that fx is not a convex combination, we have that fx = ^ E^igsucc{A) /m' where {/^}^gsucc(A) are 5i-admissible and by succ(A) we denote the immediate successors of A in A. HEREDITARY INVARIANT SUBSPACE PROPERTY 9 Remark 2.4. In a simpler manner, for any / G (see Remark ll.4l (ii)). the tree analysis of / is defined. Proposition 2.5. Let x = J2keF ^k^k be a (n,e) basic s.c.c. and G C F. Then the following holds. fcGG fcGG Proof. Let / G W^. We may assume that supp/ C G. Set Gi = {/c G supp/ : |/(efe)| ^ ^},G2 = supp/\Gi. Then clearly \Gif{Y.keG^kek)\ ^ We will show by induction that G2 G Let {/aIagA be a tree analysis of G2/. Then it is easy to see that h{A) ^ n — 1. For A a maximal node in A, we have that supp /a G Sq. Assume that for any AgA, |A| = /c>0 we have that supp /a G Sn-i-k and let A G A, such that |A| = k — 1. Then fx = ^ T!j=i /Aj> where |Aj| = k, supp/A^, G Sn-k-i for j = 1, . . . ,d and {minsupp/Aj. : j = 1, . . . , d} G 5i. Then supp /a = U^'^i supp /a^. G '5n-l-(fc-l)- The induction is complete and it follows that G2 = suppG2/ G Sn^i and therefore G2f{Y,k(iG^k(ik) ^ EfceGa < ^- Hence, |/(EfcGG CfcCfc)! < 2^ Z^fceG '^fe + ^• □ Proposition 2.6 (Basic Inequality). Let {xk}k be a block sequence in Xjgp such that \\xk\\ ^ 1, for all k and let / G W. Set (;^(/c) = maxsuppj;fc, for all k. Then there exists g G 1^(t,||||||) (see Remark 1 1.6p such that 2(7(e^(fc)) ^ /(xfc), for all /c. Proof. Let {/aIasA be a tree analysis of /. We will inductively construct {^aIagA such that for any A G A the following are satisfied. (i) gx G and 2gx{e^(j,)) > fx{xk), for any k. (ii) supp^A C {(pik) : ran fx n ranx^ 7^ 0} For A G A a maximal node, if there exists k such that ran /a Pi ran 7^ 0, set 5(A = Otherwise set (^a = 0. Let A G A be a non-maximal node, and suppose that {5'^}^>a have been chosen. We distinguish two cases. Case 1: fx is a convex combination (i.e. fx is not of type 0, Iq,, I/3, or II). If fx = I]^gsucc(A) C/i/^, set gx = Yj^i&\icc(\) '^fidfi- Case 2: fx is not a convex combination. 10 S.A. ARGYROS, P. MOTAKIS If /a = ^ Ej=i f^^J^ where succ(A) = such that f^i,<■■■<ff,a^ set Gx = {fc : ran Z;^ n ranxfc 7^ 0} Gi = {k €z G\ : there exists at most one j with lanf^. D ran j;^ ^ 0} G2 = {fc € Gx : there exist at least two j with ran f^. ranxfe 7^ 0} = {k £ Gi : ran Xfc Pi ran /^^. 7^ 0} for j = 1, . . . , d. Observe that #^2 ^ d - 1. For i = 1, . . . , d set 5^. = g^^ and for A; G G2 set gk = e*^^^y It is easy to check that if we set gx = ^{ Yl'j=i 9j + YlkeG2 ' ^^^^ S'A is the desired functional. The induction is complete. Set g = g^ □ Remark 2.7. In the previous constructions (see [3], [4], [7], [8]), the basic inequality is used for estimating the norm of linear combinations of the basis as well as to determine whether particular sequences are RIS. In the present paper the basic inequality has a weaker role, namely only to estimate the norm of the (n, e) s.c.c. In order to determine if a sequence is RIS, different techniques will be deployed in the next sections. Corollary 2.8. Let {xk}^ be a block sequence in Xjgp such that ||a;fc|| ^ 1) {cfcjfc C M and (/>(/c) = maxsuppx^ for all k. Then: II ^CfcXfcll ^ 6|| y^Cfce^(fc)||T k k Proof. Let f £ W. Apply the basic inequality and take g G ^(t,||| |||)) such that if (/'(fc) = maxsuppxfc and = sgn(ck)xk for all k, we have that '^g{e<f,(k)) ^ fiVk), for any k. It follows that k k Therefore, applying ([T|), we get II ^Cfcj;fc|| ^ 2||| ^ |cfc|e0(fc)||| = 2||| ^ Cfce<^(fc) ||| ^ 2 • 3|| ^ Cfce^(fc) ||r k k k k □ Corollary 2.9. Let x = J2T=i ^k^k be a (n, e) s.c.c. in Xjgp , such that ||xfc|| ^ 1, for k = 1, . . . ,m. If F C {1, . . . , m}, then II ^ CfcXfcll ^ ^^Ck + 12e. k£F keF In particular, we have that ||x|| ^ + 12e. HEREDITARY INVARIANT SUBSPACE PROPERTY 11 Proof. Set (j){k) = maxsuppx^, V'(fe) = minsuppxfc. Corollary 12.81 yields that II EfcGFCfcajfell ^ 6|| X;fceFCfce0(fc)||T• Since, according to the assumption, YlikeF '^k^i^ik) is a (n,e) basic s.c.c, it easily follows that Ylk&F ^ke4>(k) is a (n, 2e) basic s.c.c. By Proposition 12.51 the result follows. □ Corollary 2.10. The basis of Xjgp is shrinking. Proof. Suppose that it is not. Then there exist x* S X*gp,||x*|| = 1, a normalized block sequence {x^j/jgN in jCjgp and 5 > 0, such that x*{xk) > S, for all /c G N. Choose n G N, such that ^ < j2 and e > 0, such that e < ^. By Proposition 11.81 there exists F a subset of N, such that x = YlkeF '^k^k is a (n, e) s.c.c. By Corollary 12.91 we have that 6 > ||a;|| ^ x*{x) > S. A contradiction, which completes the proof. □ Proposition 2.11. The basis of Xjgp is boundedly complete. Proof. Assume that it is not. Then there exist e > and {xfcjfcgN a block sequence in 3£jgp, such that ||xfc|| > e and || ^^^fcll ^ 1; for all m G N. Choose ko such that d = minsuppxjtQ > |. Set Fi = {ko} and inductively choose Fi, . . . , Fd, intervals of the naturals such that (i) max Fj + 1 = minFj+i, for j < d and (ii) #Fj > max{#Fj_i, (maxsuppXmaxF,_i)^}, for I < j ^ d. Then, if we set yj = Ylik&Fj ^k, we have that || Yl'j=i VjW ^ 1- On the other hand, notice that for j = 1,. . . ,d, there exists aj an a- average in W, such that (i) ranoj C rant/j, therefore {oj}^^^ is 5i -admissible. (ii) s{aj) = #Fj, therefore {aj}j^i is very fast growing. (iii) ajivj) > e From the above it follows / = ^ Sj=i i® ^ functional of type in W and /(X]j=i Uj) > ^ > 1- Since this cannot be the case, the proof is complete. □ These last two results and a well known result due to R. C. James [17], allow us to conclude the following. Corollary 2.12. The space jCjgp is reflexive. 3. The q,/3 indices To each block sequence we will associate two indices related to a and /3 averages. In this section we will show that every normalized block sequence {xn}n has a further normalized block sequence {yn}n such that on it both 12 S.A. ARGYROS, R MOTAKIS indices a and /3 are equal to zero. As we will show in the next section, this is sufficient, for a sequence to have a subsequence generating a cq spreading model. Definition 3.1. Let {xfcjfcgis^ be a block sequence in Xjgp that satisfies the following. For any n S N, for any very fast growing sequence {agjggpj of a-averages in W and for any {-Ffej/cgN increasing sequence of subsets of the naturals, such that {aq\q<^Fu is >Sn-admissible, the following holds. For any {a^nfcjfceN subsequence of {xfcjfceN, wc have that linife X^^eFfe \(^q{^nu)\ = 0- Then we say that the a-index of {xftlfegp^ is zero and write Q;({a;fe}fe) = 0. Otherwise we write > 0. Definition 3.2. Let {xjtjfcgN be a block sequence in Xjgp that satisfies the following. For any n G N, for any very fast growing sequence {|3q\q^^ of /3-averages in W and for any {-PfejfegN increasing sequence of subsets of the naturals, such that {liq}qeFk is <Sn-admissible, the following holds. For any {xnu}km subsequence of {xfcjfcgN, we have that linifc Y^q^^Pu \P<ii^rik)\ = 0- Then we say that the /3-index of {x/jj^gf^ is zero and write fi(^{xk)k) = 0- Otherwise we write /3({a;fc}fc) > 0. Proposition 3.3. Let {xfcjfegH be a block sequence in Xjgp. Then the following assertions are equivalent. (i) a{{xk}k) = (ii) For any e > there exists j'o G N such that for any j ^ Jq there exists kj G N such that for any k ^ kj, and for any {o(q}q=i ^j-admissible and very fast growing sequence of a-averages such that s{aq) > jo, for g = 1, . . . , d, we have that Ylq=i \'^q{xk)\ < Proof. It is easy to prove that (i) follows from (ii), therefore we shall only prove the inverse. Suppose that (i) is true and (ii) is not. Then there exists e > such that for any jo € there exists j ^ jo, such that for any k^ G N, there exists k ^ kg and {aq}^^i a 5j-admissible and very fast growing sequence of a-averages with s{aq) > jo, for g = 1, . . . , d, such that X]g=i \aq{xk)\ ^ We will inductively choose a subsequence {a;„.}igN and {a^jjgN a very fast growing sequence of a-averages, such that |a*(x„.)| > |, for any i. This evidently yields a contradiction. For jo = 1, there exists ji ^ 1, such that there exists a subsequence {xkj}j€N of {ajfejfceN) a sequence {agjggN of a-averages with s{aq) > 1 for all g G N and {FjjjgN a sequence of increasing intervals of the naturals, such that: (i) {aq}q^Fj is very fast growing and ^j^-admissible. (ii) EgeF,- l«<?(a^fej)l ^ ^• (iii) If Fj = Fj \ {minFj}, then {ctq} g^u jF'. is very fast growing. HEREDITARY INVARIANT SUBSPACE PROPERTY 13 Since a[{xk}k) = 0, we have that hnij J2qeF' \'^g{^kj)\ = 0. Choose j such that \aramFjixk^)\ > f and set ni = kj,a^ = aminFy Suppose that we have chosen ni < ■ ■ ■ < Up and {a^}^_i a very fast growing sequence of a-averages, such that [a*(x„J| > |, for i = 1, . . . ,p. Set jo = max{s(aP), (maxsupp q^)^} and repeat the first inductive step to find an a-average a with s(a) > jo and Xk > Xup, Xk > a^, such that ^ |. Set Xn^+i = Xk and = a|ranzfc- The inductive construction is complete and so is the proof. □ The proof of the next proposition is identical to the proof of the previous one. Proposition 3.4. Let {xfej^gN be a block sequence in Xjgp. Then the following assertions are equivalent. (i) P{{xk}k)=0 (ii) For any e > there exists jo € N such that for any j ^ jo there exists kj € N such that for any k ^ kj, and for any {(3q}'^^^ 5j-admissible and very fast growing sequence of /3-averages such that > jo, for g = 1, . . . , d, we have that \/3q{xk)\ < S- Proposition 3.5. Let {xfcjfcgi^ be a seminormalized block sequence in Xjgp , such that either a[{xk}k) > 0, or f3[{xk}k) > 0. Then there exists c > and a subsequence {xnf,}k£N of {xfcjfcgN, that generates an spreading model, with a lower constant for all n € N. In particular, for any /co,^ € N, e > 0, there exists F a finite subset of N with miuF ^ k^ and {ck}keF, such that x = Ylik&F '^kXn,, is a (n,e) s.c.c. and ||a;|| > Proof. Assume that a{{xk}k) > 0. Then there exist £ € N,e > 0, {oglggN a very fast growing sequence of a-averages, {i<fe}fcgN increasing subsets of the naturals such that {aq}q^Fk is 5^-admissible for all A; E N and {xnj.}fceN a subsequence of {a^fcjfcgN, such that Y^qeF^ l'^g(^nfc)l > ^> for all G N. Pass, if necessary ,to a subsequence, again denoted by {xn,^}keNj generating some spreading model. By changing the signs and restricting the ranges of the aq, we may assume that X^qgiT'j. ctq(x„j,) > e, for all fc G N and rana^ C ranx^j, for any q £ Fk and A: G N. Set c = ^. Let fco,?! G N,e > 0. By Proposition 11.81 there exists F a finite subset of {uk : k^ ko} and {ck}keF, such that x = YlkaF^kXri^ is a (n,e) s.c.c. Set / = 2?T^r Z^fceF X^qeKj, '^q- Then / is a functional of type Iq in W and f{x) > 27qr7r = Arguing in the same way, for any n G N, for any F G 5„, for any {ck}k&F C M, we have that || ^keF'^kXriJ > #" T^keF kfcl- The proof is exactly the same if P[{xk}k) > 0. □ 14 S.A. ARGYROS, P. MOTAKIS Block sequences with a-index zero. In this subsection we show that seminormalized sequences with Xk = 2'^'=?/^, with {nk,ek) s.c.c. have a-index zero. Also we introduce the a-RIS sequences and we prove that the aforementioned sequences have a-RIS subsequences. Lemma 3.6. Let x = X^^i CkXk be a (n, e) s.c.c. in Xjgp , such that ^ 1, for k = 1, . . . ,m. Let also a be an a-average and set Ga = {k : ran a n r&nxk 7^ 0}- Then the following holds. \a{x)\ < min{— ^ ^ Cfc, ^ Ck + I2e} + 2max{cfc : k G Ga} Proof. Ifa = iEti/j- Set El = {k £ Ga ■ there exists at most one j with van fj n ranx^ / 0} E2 = {!,..., m}\^i Jk = {j ■ ran/j fl ranx^ ^ 0} for A; € £'2. Then it is easy to see that (2) \a{Y,CkXk)\^lYl P Ck Moreover (3) \a{ ^ CkXk)\ < 2max{cfc : k G Ga} To see this, notice that |a( ^ CkXk)\ ^ - Ck{ ^ \fj{xk)\) < max{cA: : k G Ga} — Set J = {j : there exists k G Ei such that ran fj n ran x^ / 0} and for j £ J set Gj = {A; € ii^i : ran/j n ranx^ 7^ 0}- Then the Gj are pairwise disjoint and Uj^jGj = Ei. For j € J, CorollaryESlyields that \ fjiY,keGj CkXk)\ ^ ^ Y^keGj Cfc + 12e. Therefore (4) |a(j;c,Xfc)K-^|/,(^c,Xfc)|^— ^ Cfc + 12e fcGEi ^ jeJ keGj ^ k&Ga Then ^ and (j!]) yield the following. 1 6 (5) |a( ^ CfcXfc)! ^ min{-— ^ c^, — — - Cfc + 12e} By summing up ([3]) and ([5]) the result follows. □ HEREDITARY INVARIANT SUBSPACE PROPERTY 15 Lemma 3.7. Let x = Y2T=i'^kXk be a {n,e) s.c.c. in Xjgp , such that \\xk\\ ^ 1, for k = l,...,m. Let also {ag}"^^^ be a very fast growing and cSj-admissible sequence of a-averages, with j < n. Then the following holds. fi 2-2" 2" y \ag{x)\ < + — + 18 • 2"e + 2^^e s[ai) mm supp X Proof. Set qi = min{g : rana^ fl ran a; 7^ 0}. For convenience assume that qi = 1. Then by Lemma 13.61 we have that (6) 2"|ai(x)| < -^ + 12-2"e + 2-2"e s(ai) Set Ji = {9 > 1 • there exists at most one k such that rana^ H ranx^ ^ 0} J2 = {q>l-q^ Ji} = {k : ranog n ranxj^ 7^ 0} for q> \. Gi = {k : there exists g € Ji with rana^ fl ranxfc 7^ 0} Then {minsuppxfc : A; € Gi} € 5^, hence X^fegGi '^^ ^ ^• It is easy to check that (7) 2" ^ 2^' • 2" II E ^'^^'^11 < 2'"^ For g G J2 ; Lemma 13.61 yields that 2" s(ag) 2"|ag(x)| < y Cfc + 2 • 2"max{cfc : A; G G"} 2" < — V Cfc + 2 • 2"cfc, mm supp X ^-^ where kg G C, such that Ck^ = maxjcfc : k G G^}. Then {minsuppxyt^ : g G J2} G 5j. By the above we conclude that 2 2" (8) 2"V|ag(x)|<— + 4-2"e ' mm supp x Summing up (l6|), ([7]) and ([8]), the desired result follows. □ Proposition 3.8. Let {x^l^gN be a block sequence in Xjgp such that Xk = 2"'° J2ieFk ^iVi satisfying the following: (i) {'^fcjfceN is a strictly increasing sequence of naturals. (ii) Sis-Ffe '^iVi ^ ('^fcjSfc) s.c.c. such that || ^ 1 for all i G F^, for all G N. (in) minsuppxfc > 4 • 2^"* and < (40 • 2^"'=)"^, for ah A; G N. Then a{{xk}k) = 0. 16 S.A. ARGYROS, P. MOTAKIS Proof. We shall make use of Proposition 13.31 Let e > and choose jo G N such that ^ < f • For j ^ jo, choose kj, such that > j and < |. For k ^ kj, Lemma 13.71 yields that if {0^}^^^ is a very fast growing and 5j-admissible sequence of a-averages and s{aq) > jo, for g = 1, . . . , d, we have that E. , 6 1 e e g=l □ Proposition 3.9. Let x = YlT=i ^kXk be a (n, e) s.c.c. in Xjgp , such that \\xk\\ ^ 1, for A; = 1, . . . , m, minsuppx > 4 • 2^" and e < (40 • 2^")^-*^. Then for any f W functional of type !„, such that 'w{f) = j < n, we have that 2"|/(x)| < ^. Proof. Let / = ^ Ylq=i ^ functional of type Iq, with weight w{f) = j < n. Then Lemma 13.71 yields that 1 1 c o on 7 2"|/(x)| ^ — (2"y |a„(x)|) < — (^^+ +18-2"e+22M ^ — i.yv;i\2J^ Z^i 9V 2J ai minsuppx ^ ^ 2J q=l □ Definition 3.10. A block sequence {xk}k is said to be a (C, {71^}^) a- rapidly increasing sequence (or {C,{nk}k) a-RIS), for a positive constant C ^ 1 and a strictly increasing sequence of naturals {n^lfc, if ||xfc|| ^ C for all k and the following conditions are satisfied. (i) For any k, for any functional / of type Iq, of weight w{f) = j < we have that \f{xk)\ < ^ (ii) For any k we have that ^nl+i maxsuppx^ < Remark 3.11. Let {x/cj/cgN be a block sequence in Xjgp , such that there exist a positive Constance C and {nk}keN strictly increasing naturals, such that \\xk\\ ^ C for all k and condition (i) from Definition 13.101 is satisfied. Then passing, if necessary, to a subsequence, {xfcjfcgN is (C, {rifclfceN) a-RIS. Proposition 3.12. Let {xk}k£N be a block sequence in Xjgp with x^ = 2"*= X^ieFfc '^iUi satisfying the following: (i) {nk}keN is a strictly increasing sequence of naturals. (ii) J^jgj? cfyf is a {nk,ek) s.c.c. such that || ^ 1 for all i G F^, for all k£N. (iii) minsupp > 4 ■ 2^"'= and Sk < (40 • 2^"'=)"^ for all A; G N. Then passing, if necessary, to a subsequence, {xfcjfcgN is (7, {nk}k) a-RIS. Proof. Corollary 12.91 yields that ||a;fc|| < 7 while Proposition 13.91 yields that (i) from Definition 13. 101 is satisfied. By Remark 13.111 the result follows. □ HEREDITARY INVARIANT SUBSPACE PROPERTY 17 Block sequences with /3-index zero. In this subsection we first prove that every increasing seminormalized (n, En) s.c.c. built on an a-RIS block sequence has /3-index zero. This yields that every block sequence has a further block sequence with both a, (3 indices equal to zero. We start with the following technical lemma. Its meaning becomes more transparent in the following Corollary 13.141 and Lemmas I3.15( 13.161 Notation. Let x = Yl'k=i^kXk be a (n,e) s.c.c. such that {xk}^^i is (C,{nfc}^=i) a-RIS with 2^" < m. Let also / = \ Yfj=ifj be a type II functional. Set Jo = {j : n 5^ ^/,) < 22"} h = {] ■■ w{fj) < n} h = {3:2'^^w{f,)<n^} Jk = {j ■■ nk ^ w{fj) < Uk+i}, for k <m and = {j : rim < w{fj)} Under the above notation the following lemma holds. Lemma 3.13. Let x = X^^Li CkXk be a (n, e) s.c.c. ^ 2 such that the following are satisfied: (i) {xk}i:=i is (C,{nfc}-^,) a-RIS with 2^- < m. (i) minsuppx > 4C • 2^'" and e < (40C23")~^ Let also f = \ Yl'j=i fj be a functional of type II. Then there exists Fj C {xk ■ ran / PI ran j;^ 7^ 0} with {minsuppj;fc ■ k £ Ff} G S2 such that ^k=2jeJk k=l jeJk + y — ^ + y + cr y ck iG/i jG/2 ^ kGFf Proof. Notice that {Jk}k^=i are disjoint intervals of {1, ... , d} and that gk = hEjejJj G VF, for fc= l,...,m. Set Ff = {k : van gk Hranxfc ^0}- It easily follows that {minsuppx^ : k G Ff} G S2 and that on Let ko ^ m, j G Jk^- Then (10) 2-\MY,CkXk)\ < C^;^^j^j^ and 2"j/,( c,x,)| < 18 S.A. ARGYROS, P. MOTAKIS Proposition 13.91 yields that for j G Ii we have that 2"|/j (x)| < -^if-j and hence For j e I2 we have that 2"|/j(x)| < and therefore je/2 je/2 Corollary 12 . 91 yields that 2"||x|l < 7C, and since Iq is an interval, it follows that i T^jeio fj Therefore nn (13) yIE/^-(^)i<7^ Summing up ([2]) to ([T3D the desired result follows. □ The next corollary will be useful in the next sections, when we define the notion of dependent sequences. Corollary 3.14. Let x = X^fcLi CfcX^ be a (n,e) s.c.c. in Xjgp,?7, ^ 3 such that the following are satisfied: (i) {xk}T=i is (C,K,}- a-RIS with 2^^ < m. (i) minsuppx > 4C • 2^" and e < (40C23")-i. Let also f = \ Sj=i fj be a functional of type II with w{f) n {n, . . . , 2^^^} = 0. Then 2'^|/(x)| < C(— + 4-+ y ^7T + 2"e) {i:«'(/j)<"} Proof. Apply Lemma 13.131 Then the following holds. (14) 2-i/wi < 2 ^fc=2jeJfe k=i jeJi, + (:72"e Notice the following. ^ 2"fc 2^ 2 (^^) H H 2«'(/j)+«fc-i ^ 2^ ^ 22" 2" T-:,^ ^-^ 2" / 1.2 (16) E:^ + EE:w7:t = 2"( E 2u'(/j) ^ 2'"(-/'j) ^ ^ 2'"(/j)'^ 2*^ HEREDITARY INVARIANT SUBSPACE PROPERTY 19 Applying ^ and (HH]) to ([HI the result follows. □ Lemma 3.15. Let x = YlT=i '^k^k be a (n, e) s.c.c. in X^^p , n ^ 2 such that the following are satisfied: (i) {xk}T=i is {C,{nk}lU) a-RIS with 2^'' < m. (ii) minsuppx > 4C22" and e < (40C23")-i. Let also /3 be a /3-average. Then there exists C {xk ■ ran ^ n ran x^ 7^ 0} with {minsuppxfc : k £ Fjj} G S2 such that Proof. If /3 = i X]g=i then by definition the /g are functionals of type II with disjoint weights w{fq)- For convenience, we may write fq = where the index sets Gg,q = 1, . . . ,p are pairwise disjoint. Notice that for ii,j2 € G,ji 7^ j2 we have that w{fj^) 7^ w[fj^). By slightly modifying the previously used notation, set G = U^^j^Gg and /o = {jeG:n^w{fj)<2'^} h = {j G G : < n} /2 = {jGG:22"^^/,)<ni} Jk = {j ^ G : Uk ^ w{fj) < Uk+i}, for k <m and Jm = {j G G : Tim ^ Wifj)} By Remark ll . 1 1 there exists at most one qo ^ d, with w{fq^^)ri {n, . . . , 2^"} 7^ and if such a go exists, then ij^w{fqQ) n {n, . . . , 2^*^} ^ 1. Apply Lemma 13.131 Then for q = 1, . . . , d there exists Fq C {xfc : ran /? R ranxfc 7^ 0} with {minsuppxfc : k £ Fq\ £ S2 such that m-l (17) 2"|/3(x)| < + +VV ^ ^ ^k=2j&Jk k=l j&Jk + > ^7T + > ^7T + -'-2" > > Cfc je/i ie/2 ^ q=\k(iF^ Just as in the proof of Corollary 13.141 notice the following. ™_ 2"fc \ (1^) X] X] 2"'(/i)+"fe-i ^ 22" 2n ^ 2" 2 20 S.A. ARGYROS, P. MOTAKIS By the definition of the coding function a we get ^ ^ 1000 J 6/1 1 ^ (21) -C2"J] J]cfe^C2"max{ j;cfc:(? = l,...,p} = C2" Ck ^ q=l k&Fq keFq ^SFq^ for some 1 ^ qo ^ p. Set = Fq^ and apply (fT8|) to (f2T]) to pT|) to derive the desired result. □ Lemma 3.16. Let x = X^^^ c^x^ be a (n, e) s.c.c. ^ 4 such that the following are satisfied: (i) {xk}T=i is (C^,K}r=i) «-I^IS with 22^^ < m. (i) minsuppx > 4C22" and e < (40C23")-i. Let also {/3g}g^]^ be a very fast growing and 5j-admissible sequence of /3-averages with j ^ n — 3. Then we have that 2" ^ < ^ + 2C2"e + C2'-e q=l q=l ^^Pl> Proof. Set Ji = {q : there exists at most one k such that ran/?^ n ranx^ / 0} J2 = {1,...,4\J2 Gi = {k : there exists q G Ji with ran /3g n ranx^ 7^ 0} Then {minsuppx^ : € Gi} € 5„_2 and it is easy to check that d (22) 2" J] |/3,(x)| ^ 2^2"|| CkXkW < C2'^e qeJi keGi For q £ J2, choose Fg C {1, . . . , m} as in Lemma . 1 5 1 and set F = Uggjj-Fg. Then {minsuppxfc : G F} G 5„_i, therefore X]geJ2 SfceF, ^fc < 2e. Lemma 13.151 yields that (23) 2"^|/3,(x)|< - + 2C72"e Combining (p2|) and ([23l) . the result follows. □ Proposition 3.17. Let {yjjjeN be {C,{ni}i) a-RIS in Xjgp, {xjtljtgN be a block sequence of {yjjigN, Xk = 2™'* X^igFfc '^iVi satisfying the following: (i) {mk}km is a strictly increasing sequence of naturals. (ii) SieFfe cfyj is a {mk,ek) s.c.c. with 2^™'= < rirainFf HEREDITARY INVARIANT SUBSPACE PROPERTY 21 (iii) minsuppj;fc > 4(722™'= and et < (40C23"^'=)-\ for all /c € N. Then a[{xk}k) = as well as P[{xk}k) = 0. Proof. Proposition 13.81 yields that a[{xk}k) = 0. To prove that P[{xk}k) = 0, we shall make use of Proposition 13.41 Let £ > and choose jo G N such that 8C ^8C £ For j ^ jo choose kj, such that rrik. ^ j + 3 and , < |. For A: ^ Lemma [3.161 vields that if {/3g}J^^^ is a very fast growing and 5j-admissible sequence of /3-averages and s(/3g) > j'o, for g = 1, . . . , d, we have that El „ , , I 8C 8C 3 £ e \Pa{xk)\ < h > ^ H <- + - = £ 9=1 J>JO □ Corollary 3.18. Let {xfoj^gN be a normalized block sequence in Xjgp . Then there exists a further normalized block sequence {y/cj/cGN of {xk}k&i-, such that a[{yk}k) = as well as = 0. Proof. If a(y{xk}k) = and I3{^{xk}k) = 0, then there is nothing to prove. Otherwise, if > or /3({a:;fc}fc) > 0, apply Proposition 13.51 to con- struct a seminormalized block sequence {zk}k&i^ satisfying the assumption of Proposition [3T8l Then a[{zk}k) = 0. Proposition [3T2] yields, that pass- ing, if necessary, to a subsequence, we have that {zk^k&n is (7, {nk}k) a-RIS. If = 0, set yk = jj^^fc and {yk}km is the desired sequence. Otherwise, if I3(y{zk}k) > 0, apply once more Proposition 13.51 to con- struct a seminormalized block sequence {i«fc}fceN) satisfying the assumption of Corollarv 13.171 Set = -n^—n Wk and {y^lfceN is the desired sequence. □ 4. cq spreading models This section is devoted to necessary conditions for a sequence {2;^}^ to generate a cq spreading model. At the beginning a Ramsey type result is proved concerning type II functionals acting on a block sequence {xfcj/c with /3({2;fc}fc) = 0. Then conditions are provided for a finite sequence to be equivalent to the basis of ^J^. This is critical for establishing the HI property and the properties of the operators in the space. Moreover it is shown that any block sequence {xk}k with a[{xk}k) = and = contains a subsequence generating a cq spreading model. Another critical property re- lated to sequences generating cq spreading models is that increasing Schreier sums of them define a-RIS sequences. 22 S.A. ARGYROS, R MOTAKIS A combinatorial result. Definition 4.1. Let xi < X2 < X3 be vectors in Xjgp, / = ^Yl'j=ifj be a functional of type II, such that supp / fl ranxj 7^ 0, for i = 1,2,3 and jo = min{j : ran/j n ran 2:2 7^ 0}- If ran/jg fl lanxs = 0, then we say that / separates xi,X2,X3. Definition 4.2. Let i,j G N. If there exists f e W a functional of type II, such that i,j G w{f), then we say that i is compatible to j. Lemma 4.3. Let xi < a;2 < • • • < Xm be vectors in Xjgp, such that there exist £ > and {fkY^''^^^ functionals of type II satisfying the following. (i) fk separates xi,Xk, x^, for A; = 2, . . . , m — 1 (ii) If fk = ^Ejii/j' and jk = mm{j : ran/j= n vanxk 7^ 0}, then w{fj^) is not compatible to w{fj^) for k ^ £. (iii) \fk{xm)\ > e for /c = 2, . . . , m - 1 Then there exists a /3-average /? in of size s(/3) = m — 2 such that P{Xm) > e. Proof. Set Qk = sgn(/fe(xm))/fc|ran2;™, for /c = 2, . . . , m - 1. Then gk is a functional of type II in W. We will show that the have disjoint weights w{gk)- Towards a contradiction, suppose that there exist k ^ £ and i G w{gk) n w{ge)- By (i) and the way type II functionals are constructed, it follows that //cl[minsuppa;2,...,™iiisuppa;m._i] /€|[minsuppa;2,...,™iiisuppa;m-i] ' '-'^■^i^ COntradictS (ii). By the above, it follows that if we set /3 = Y1T=2 9kj then /3 is the desired /3-average. □ Lemma 4.4. Let xi < a;2 < • • • < Xm be vectors in Xjgp , such that there exist e > and {fk}k'=2 functionals of type II satisfying the following. (i) fk separates xi,Xk, Xm, for fc = 2, . . . , m — 1 (ii) If fk = \ Y.f=ifj and jk = min{j : ran/j= n ranxfc ^ 0}, then ^^^{fl) = w{f'^,)^oxki^L (iii) If i'f^ = min{j : ran ("Iran ^ 0}, then w{f^, ) ^ w{fi) for k ^ £. (iv) \fk{xm)\ > £ for = 2, . . . ,m - 1 Then there exists a /3-average /3 in W of size s{j3) = m — 2 such that /3{xm) > e. Proof. As before, set gk = sgn{fkixm))fk\Ta.nxm^ for /c = 2, . . . , m - 1. Then gk is a functional of type II in W. We will show that the gk have disjoint weights w{gk). Suppose that there exist k ^ i and i G w{gk) n w{gi). By (i), (ii) and the way type II functionals are constructed, it follows that /fe|[minsuppa;2,.--iniiiisuppa;m] ~ |[minsuppx2,...,minsuppa;m] HEREDITARY INVARIANT SUBSPACE PROPERTY 23 This leaves us no choice, but to conclude that w{fj, ) = w{fj,), a contradic- tion. It follows that if we set /3 = Y1T=2 9k ^ then (3 is the desired /3-average. □ Proposition 4.5. Let {xfcjfceN be a bounded block sequence in Xjgp , such that /3({xfc}fc) = 0. Then for any e > 0, there exists M an infinite subset of the naturals, such that for any ki < k2 < (z M, for any functional f GW of type II that separates x^^ , , , we have that |/(x/c-)| < e, for some {1,2,3}. Proof. Suppose that this is not the case. Then by using Ramsey theorem [22], we may assume that there exists e > such that for any k < i < m £ N, we have that there exists fk,e,m a functional of type II, that separates and \fk,e,m{xk)\ > e, \fk,e,m{xe)\ > e and \ fk,i,m{xm)\ > e. For 1< < m, if = i Y:%i set ik,m = min{i : lanfj''^ n ranxi ^ 0} jk,m = min{j : ran /j"'™ n ran Xfc / 0} jk,m = min{j : ran /j"'™ n ran / 0} Notice, that for 1 < < m, since \ fi,k,mixi)\ > £, it follows that 1 £ > 2'^(Afe™ ) II max supp xi By applying Ramsey theorem once more, we may assume that there exists ni G N, such that for any 1 < A; < m, we have that w{f^^^) = ni Arguing in the same way and diagonalizing, we may assume that for any A; > 1, there exists € N such that for any m > k, we have that Ai = {{kj} -.nk^ui and Uk is compatible to n^} A2 = {{k,£} G [N \ {1}]^ : Uk n£ and is not compatible to n^j ^3 = {{k,e} ■.nk = ne} Once more, Ramsey theorem yields that there exists M an infinite subset of the naturals, such that [M]^ c Ai,[M]^ C A2, or [Af]^ c ^3. Assume that [M]-^ C Ai and for convenience assume that M = N \ {!}. Choose ko > 1 such that /cq > max supp xi. Since ni is compatible to n2 and in general nk~i is compatible to n^, for A; > 1, it follows that there exists a functional f = \ Sj=i fj of tyP^ II in W , such that ran / fl ranxi ^ and for A: = 1, . . . , Aq there exists j^, with w{fj^) = Uk, for A; = 1, ... , kQ. Since minsupp/i ^ max supp xi it follows that {fj}j=i can not be 5i- admissible, a contradiction. 24 S.A. ARGYROS, P. MOTAKIS Assume next that [M]^ C ^2- Lemma HT3l yields that > and since this cannot be, we conclude that [M]^ C A3, therefore there exists no € N, such that = no, for any k £ M. Assume once more that M = N \ {1} and set B = {{k,i,m} e [n\{i}f : wifp'n = wifyn} If there exists M an infinite subset of the naturals, such that [M]^ C B'^, Lemma [4.41 yields that P[{xk}k) > 0, therefore by one last Ramsey argument, there exists M an infinite subset of the naturals, such that [M]^ C B. By the above, we conclude that for m ^ 4, ranx^ C ran/|j™ and \f]Z(^k)\>2e,foik = 2,...,m-2. Set fm = f'j^ ^-nd let / be a li;* limit of some subsequence of {fm}m&i- Then ^ 2e, for any k^ 2. Corollary 12.101 yields a contradiction and this completes the proof. □ Remark 4.6. The proof of Proposition 14.51 is the only place where the condition I3(^{xk}k) = is needed. This makes necessary to introduce the /3-ayerages and their use in the definition of the norm. Finite sequences equivalent to i^o basis. Proposition 4.7. Let xi < ■ ■ ■ < x„ be a seminormalized block sequence in jCjgp , such that \\xk\\ ^ 1 for A; = 1, . . . , n and there exist n + 3 ^ ji < ' ' ' < in strictly increasing naturals such that the following are satisfied. (i) For any k^ G {l,...,n}, for any k ^ k^^k G {l,...,n}, for any {ag}q^2 fs-st growing and 5j-admissible sequence of a-ayerages, with j < jfcg and s{ai) > minsuppa^fc^, we have that Ylq=i |q^<?(^A;)I < 1 n-2" • (ii) For any /cq G {l,...,n}, for any k ^ kQ,k {l,...,n}, for any {f3q}q^i yery fast growing and 5j- admissible sequence of /3-ayerages, with j < and s(/3i) > minsuppx^p, we have that Ylq=i \(^q(.^k)\ < 1 n-2" ■ (iii) For A; = l,...,n— 1, the following holds: maxsuppxfc < (iy) For any 1 ^ ki < k2 < k^ ^ n, for any functional f £ W of type II that separates Xki,Xk2,Xk^, we have that \f{xkj\ < for some iG {1,2,3}. Then {xk}^^i is equivalent to £^0 basis, with an upper Constance 3 + Moreover, for any functional / G of type Iq, with weight = j < ji, we have that \ f{Ylk=i^k)\ < Proof. By using Remark ll.3l we will inductively prove, that for any {ck}^^i C [— 1, 1] the following hold. HEREDITARY INVARIANT SUBSPACE PROPERTY 25 (i) For any f eW, we have that |/(I]fc=i CfcXfc)| < (3 + Jr) max{|cA:| : = 1, . . . , n}. (ii) If / is of type la and w{f) ^ 2, then \f(J2k=i'^kXk)\ < (1 + Jr)niax{|cfc| : k = 1, . . . ,n}. (hi) If / is of type I« and w{f) = j < ji, then \fiJ^k=i(^kXk)\ < ^"^j" max{|cfc| : fc = 1, . . . , n}. For any functional / G Wq the inductive assumption holds. Assume that it holds for any / G Wm and let / € Wm+i- If / is a convex combination, then there is nothing to prove. Assume that / is of type Iq, / = i Ylq=i '^qi where {0^}^^]^ is a very fast growing and cSj-admissible sequence of a-averages in Wm- Set ki = min{A; : ran / n ran Xfc 7^ 0} and qi = min{g : ran Uq fl ran 7^ 0}. We distinguish 3 cases. Case 1: j < ji- For q > qi, we have that s{aq) > minsuppxjt^, therefore we conclude that n ^ (24) ^ \agC^CkXk)\ < — max{|cfc| : k = l,...,n} q>qi k=l while the inductive assumption yields that n 2 (25) \aq^C^CkXk)\ < (3 + — )max{|cfe| :k = l,...,n} k=l Then (I24p and (|25p allow us to conclude that " 3 + 4r (26) \f(^CkXk)\ < max{|cfc| ■.k = l,...,n} k=l Hence, (iii) from the inductive assumption is satisfied. Case 2: There exists ko < n, such that jk^ ^ j < jko+i- Arguing as previously we get that 3 + 4r (27) \f(.y~] CkXk)\ < . ^" max{|cfc| : k = 1,. . . ,n} k>ko and (28) \fi^ CkXk)\ < ^max{|cfc| :k=l,...,n} k<ko Using (i27|l . (f28]l . the fact that \f{xko)\ ^ 1 and jki^ ^ n + 3, we conclude that n 2 (29) \fC^CkXk)\ < (1 + — )max{|cfc| :k = l,...,n} k=i ^ Case 3: j ^ jn 26 S.A. ARGYROS, P. MOTAKIS By using the same arguments, we conclude that n ^ (30) \fC^CkXk)\ < (1 + — )max{|cfc| :k = l,...,n} k=l Then (j26|) . ([29|) and (f30|) yield that (ii) from the inductive assumption is satisfied. If / is of type I^, then the proof is exactly the same, therefore assume that / is of type II, / = ^ Sj=i /?> where {fj}j^i is an 5i-admissible sequence of functionals of type Iq, in Wm- Set E = {k:\f{xk)\^^} n ■ 2" El = {k £ E : there exist at least two j such that ran/j n ranx^ ^} Then #Ei ^ 2. Indeed, if ki < k2 < k^ £ Ei, then / separates Xk^,Xk2 and Xfeg which contradicts our initial assumptions. If moreover we set J = {j ■ there exists k G E \ Ei such that ran fj ranxfc 7^ 0}, then for the same reasons we get that #J ^ 2. Since for any j, we have that w{fj) G L, we get that w^fj) > 2, therefore: n 2 (31) |/( X] < (l + ^)max{|cfc| : A;= l,...,n} n, (32) |/(^CfcXfc)| ^ 2max{|cfc| : A; = 1, . . . ,n} n ^ (33) |/(^Cfcj:fc)| ^ n • — ^max{|cfc| : A; = 1, . . . ,n} Finally, ([M]) to ([MD yield the following. n 2 |/(^CfeXA..)I < (3 + — )max{jcA:| : A; = 1, . . . ,n} fc=i This means that (i) from the inductive assumption is satisfied an this com- pletes the proof. □ The spreading models of Xjgp . In this subsection we show that every seminormalized block sequence has a subsequence which generates either ii or Co as a spreading model. Proposition 4.8. Let {x^l^gN be a seminormalized block sequence in Xjgp , such that \\xk\\ ^ 1 for all A; € N and a[{xk}k) = as well as P[{xk}k) = 0. Then it has a subsequence, again denoted by {xfcjfcgN satisfying the following. (i) {xfcjfcgN generates a cq spreading model. More precisely, for any n ^ A:i < • • • < A;„, we have that || Y17=i \\ ^ 4. HEREDITARY INVARIANT SUBSPACE PROPERTY 27 (i) There exists a strictly increasing sequence of naturals {jnlneNj such that for any n ^ ki < ■ ■ ■ < kn, for any functional / of type with ^if) = j < jn, we have that " 4 \fC^XkJ\ < — 1=1 Proof. By repeatedly applying Proposition 14.51 and diagonalizing, we may assume that for any n ^ ki < k2 < k^, for any functional / of type II that separates Xj.^,Xk2 ^^'^ ^fcs' '^^ have that \f{xk.)\ < i^^-^rr, for some {1,2,3}. Use Propositions 13.31 and 13.41 to inductively choose a subsequence of {xk}k£N, again denoted by {xfcjfceN and {jfc}fcgN a strictly increasing se- quence of naturals with jk ^ k + 3 for all /c G N, such that the following are satisfied. (i) For any ko G N, for any k ^ kQ, for any {aq}'^^i very fast growing and 5j-admissible sequence of a-averages, with j < j^p and s{ai) > minsuppxfco, we have that Y,'!i=i \aq{xk)\ < (ii) For any k^ G N, for any k ^ k^, for any {/3q}^^i very fast growing and tSj -admissible sequence of /3-averages, with j < jk^ and > minsuppj;fco, we have that X]g=i \Pqixk)\ < (iii) For G N, the following holds: max supp < It is easy to check that for n ^ ki < ■ ■ ■ < kn, the assumptions of Proposition 14.71 are satisfied. □ Propositions 13.51 and 14.81 yield the following. Corollary 4.9. Let {xfcjfcgN be a normalized weakly null sequence in Xjgp. Then it has a subsequence that generates a spreading model which is either equivalent to cq, or to £i. Proposition 4.10. Let {xjtjjtgi^ be a normalized block sequence in Xjgp, that generates a cq spreading model. Then there exists {-FfcjfceN an increas- ing sequence of subsets of the naturals such that ^ minF^ for all A; G N and lim^ = oo such that by setting = X^jg^? x^, there exists a subsequence of {yk}k<^N, which generates an ^" spreading model, for all n G N. In particular, for any ko,n G N and e > 0, there exists F a finite subset of N with minF ^ /cq and {ck}kGF, such that (i) X = J2keF^kyk is a (n,e) s.c.c. (ii) ^ < ||x|| ^ f (iii) {vkjkeF is {'i,{nk}kGF) a-RIS and 2^" < UminF- (iv) For any 77 > there exists a functional / of type Iq, of weight w{f) = n such that f{x) > and maxsupp/ > maxsuppx. 28 S.A. ARGYROS, P. MOTAKIS Proof. Since generates a cq spreading model, Proposition 13 . 51 yields that a[{xk}k) = as well as I3[{xk}k) = 0, therefore passing, if necessary, to a subsequence {xk}keN, satisfies the conclusion of Proposition 14.81 Choose {-FfcjfcgN an increasing sequence of subsets of the naturals, such that the following are satisfied. (i) #Fk < rniuFk for ah A; G N. (ii) > max{#Ffc, (maxsuppXmaxFj^}, for all A; G N. By Proposition 14.51 and Remark 13. IH we have that 1 ^ WukW ^ 4, for all k €N and passing, if necessary, to a subsequence, {yk}keN is (4, {nk}keN) RIS. Moreover it is easy to see, that for any /c € N, r/ > 0, there exists an a-average a of size s{a) = #Fk, such that a{yk) > 1 — ry and ran a C yk- This yields that a(^{yk}k) > 0, therefore we may apply Proposition 13.51 to conclude that {yk}keN has a subsequence generating an spreading model, for all n G N. We now prove the second assertion. Let ko,n G N and e > 0. By taking a larger /cg, we may assume that n^Q > 2^". Also, by taking a smaller e, we may assume that e < (160 • 2^")^^. Set e' = e(l — e) Proposition 11.81 yields that there exists {di, . . . ,dm} a finite subset of {fc : /c ^ ko} and {ck}]^^i such that x' = YlT=i ^kydk is a (n, e') s.c.c. It is straightforward to check that x = Xlfc^i i-c V^k ^ ('^j ^) s.c.c. By Corollarv 12.91 we have that ||x|| ^ ||. For some r/ > 0, /c = 1, . . . , m, choose an a-average ak of size s(afc) = #Frf^, such that UkiydJ > 1 - f? and rana C yk- Set / = ^(X^^^iafc), which is a functional of type Iq of weight w{f) = n such that f{x) > and maxsupp/ > maxsuppx. □ Corollary 4.11. Let 1" be an infinite dimensional closed subspace of Xjgp. Then Y admits a spreading model equivalent to cq as well as a spreading model equivalent to ii. Proof. Assume first that Y is generated by some normalized block sequence {x/fc}/fcgN- Corollary 13.181 and Proposition 14.81 yield that it has a further normalized block sequence {yfcjfceNj generating a spreading model equivalent to Cq. Proposition 14.101 yields that {yfcjfceN has a further block sequence gener- ating an ii spreading model. Since any subspace contains a sequence arbitrarily close to a block se- quence, the result follows. □ HEREDITARY INVARIANT SUBSPACE PROPERTY 29 We remind that, as Propositions 13.51 and 14.81 state, if a sequence generates an li spreading model, then passing, if necessary, to a subsequence, it gener- ates an (.\ spreading model for any A; € N. However, as the next proposition states, the space Xjgp does not admit higher order cq spreading models. Proposition 4.12. The space Xjgp does not admit Cq spreading models. Proof. Towards a contradiction, assume that there is a sequence {x^j/jgN hi Xjgp , generating a Cq spreading model. Then it must be weakly null and we may assume that it is a normalized block sequence. By Proposition 14.101 it follows that there exist {-Ffcj/cgN increasing, Schreier admissible subsets of the naturals and c > such that || X]j=i SigF^ ^ n ■ c for any n ^ A;i < . . . < Since for any such F]^^ < ■ ■ ■ < Fj.^ we have that U^^iF^^ € 52, it follows that {a^fcjfceN does not generate a Cg spreading model. □ Corollary 4.13. Let Y be an infinite dimensional closed subspace of Xjgp. Then Y* admits a spreading model equivalent to £i as well as a spreading model equivalent to Cg, for any n € N. Proof. Since Y contains a sequence {xfcjfcgN generating a spreading model equivalent to cq, which we may assume is Schauder basic, then for any normalized {a^^jfcgN C Y*, such that x^(xm) = 6n,m for n, m G N, we have that passing, if necessary, to a subsequence, {x^jfcgN generates a spreading model equivalent to £i. To see that Y* admits a spreading model equivalent to Cq for any n G N, take the previously used sequence {x^j^gN- Working just like in the proof of Proposition 14. 10] find {i^fcjfcgN successive subsets of the natural such that minFfc ^ #Ffc, for ah fc S N, if = J2ieF^: A; G N, then {yk}k€N is seminormalized and there exists a very fast growing sequence of a-averages {a/cjfceN C W such that lim'mf Uki^^^p^ Xi) ^ 1. Then, if c = limsup;. ||yfc||,we evidently have that liminffc ||aA;|| ^ 1/c and since for any n G N, F G 5„, we have that -^Ylfq^F '^q ^ functional of type Iq, in W, it follows that || Ylfq^F^qW ^ This means that, {ak\ken generates a spreading model equivalent to Cq, with an upper constant 2". Let /* : X*gp — > Y* be the dual operator of I : y — > Xjgp . Then {I*{ak)}ken generates a spreading model equivalent to Cg, for any n G N. Since ||/*j| = 1, all that needs to be shown is that liminffc > 0. Indeed, liminf ||r(afc)|| ^ lim inf (/* a^) ( ) = lim inf ( ^'^^'^ ) ^ 1/c □ 30 S.A. ARGYROS, P. MOTAKIS 5. Properties of Xj^p and C{X^^p) In this final section it is proved that j£jgp is hereditarily indecomposable and the properties of the operators acting on infinite dimensional closed subspaces of Xjgp are presented. Dependent sequences and the HI property of Xjgp . In the first part of this subsection we introduce the dependent sequences, which are the main tool for proving the HI property of Xjgp and studying the structure of the operators. Definition 5.1. A sequence of pairs {{xk, fk)}k=i^ where xi < • • • < x„ G Xjgp and /i < • • • < /n £ W, is said to be a 1-dependent sequence (respec- tively a 0-dependent sequence) if the following are satisfied. (i) {fk}k=i is 5i-admissible special sequence of type 1^ functionals, fkixk) = 1 when {{xk,fk)}k=i is 1-dependent (respectively fkixk) = when {{xk, fk)}k=i is 0-dependent), maxsuppx^ < maxsupp fk for k = 1, . . . ,n and ran fk fl ran Xm = for k ^ m. (ii) Xk = 6'fe2™'^ XliGFfe CiVi, where T^ieF^ '^iVi is a (mfc,efc) s.c.c. with rrik = w{fk) and ^ ^ 6*^ < ||, for /c = 1, . . . , n. (iii) {y^}i(zF^^ is (4, {njigi?^) a-RIS with 2^'"'= < '^min Fl- ■ (iv) minsuppxfc > 240n2"22'"* and Sk < (320n2"23™''-)-^ (v) 2™i > 60n2'^ and if po = minsuppxi, then ^ + Ep>po ^ < Proposition 5.2. Let {{xk, fk)}'kLi be a 1-dependent sequence in Xjgp and set Uk = X2k-i — X2k, for k = \, . . . ,n. Then we have that: (i) 'n\\T.'^=l^k\\^l (ii) ^IIELi^^lK^ Proof. Since | Y^h=i fk is a type II functional in W , it immediately follows that i||I]fc"iXfc|| ^ ^Yi=ifk{xk) = 1- By Corollary ESI it follows that 1 = fk{xk) ^ \\xk\\ ^ 7 • 46*^ ^ 29 and this yields that 1 ^ ||yfc|| ^ 58, for /c = 1, • • • , n. Set jk = 'ni2k-i — 2. We will show that the assumptions of Proposition 14.71 are satisfied. Prom this, it will follow that ^|| Yllt=i VkW ^ ^8^, which is the desired result. The first and second assumptions, follow from Lemmas 13.71 and 13.161 re- spectively and the definition of the 1-dependent sequence. The third assumption follows from the fact that, by the definition of the 1-dependent sequence, maxsupp/^ > maxsuppx^, for k = l,...,2n and the definition of the coding function a. It remains to be proven that the fourth assumption is also satisfied. Let 1 ^ fci < A;2 < A;3 ^ n and g = ^ Sj=i 9j be a functional of type II that separates y^^ , yk^ and yk^ ■ Set jo = min{j : langj fl ran 2/^3 7^ 0} and assume first that 'w{fjg) = m2fc3-i Since supp 5 n supp y^^ / 0, it follows that gj^-i = f2k-i-2 and there HEREDITARY INVARIANT SUBSPACE PROPERTY 31 exists / an interval of the naturals, ran 1/^2 ^ ^1 such that g = Ylk=i fk)- This yields that ^(yfcj) = 0- Otherwise, if 'w{fjg) ^ rn2ks-i, set g' = g\ra.nyk.^ and Corollarv 13.141 yields the following. \9'{yks)\ < 2 • ^^(^mzfcg-i + 22m2fc3_i + 2"'fe) ^'"^''^~^^"'2'=3-i) ie«;(g'): ■!«(3j)<n Since (7 separates yk-^ , and y^g , we have that min w{g') ^ po = min supp xi , therefore 1 ^ 1 1 1 ^ 2"'fe) ^ 2^ 2P ^ ^ p2 < 280 ■ 2n22" jeiS{g'): P^PO P^Po w{gj)<n Moreover, we have that 11^ 3 1 2m2fc3-i "T 22m2fc3_i ^m2fe3-i ^ 20 • 2n22" We conclude that [^(yfcg)! < 2n2'^" ^ nl"' "^liich means that the fourth assumption is satisfied. □ The next proposition is proved by using similar arguments. Proposition 5.3. Let {{xk^ fk)}'k=i ^ 0-dependent sequence in Xjgp. Then we have that: InV^ „ 112 -II > XkW ^ n ^-^ n k=l We pass to the main structural property of Xjgp . Theorem 5.4. The space Xjgp is hereditarily indecomposable. Proof. It is enough to show that for X, Y block subspaces of Xjgp , for any e > 0, there exist x G X and y (^Y, such that ||x + y|| ^ 1 and ||x — y|| < e. Let n G N, such that ^ < e. By Corollarv 13.181 and Proposition 14.81 there exist {xfcjfceN a normalized block sequence in X and {yk}k€N a normalized block sequence in Y, both generating cq spreading models. Choose m-i G Li (see the definition of the coding function) such that 2-1 > 60 • 2n22", po G N such that ^ + Ep>po ^ < 280.2«22^22-i ^nd < ei < (320 • 2n2"23'"i)-i. By Proposition 14.101 there exists x'/ = J2ieFi ^ ("^ij^i) s.c.c. and /i a functional of type Iq such that I'i'l < \\r"\\ < 2™! ^ 11 1 II ^ 2™! (h) {xk}keFi is (4, {nfclfcgFj a-RIS and 2'^'^^ < rimmF- (iii) /i is of weight w{fi) = mi and fi{x'() > 29^?m" ^'^d maxsupp/ > max supp x'(. 32 S.A. ARGYROS, P. MOTAKIS Set 01 = (2™i/i(4))"' and x'^ = 9^2^^ EieF, Continue in the same manner and construct x'^ , . . . , and /i , . . . , f2n such that {(a^fc, /fc)}fc=i is a 1-dependent sequence and x'2j^_i G ^■,^'2k ^ for fc = 1, . . . , n. Set X = i XlLi and y = ^ Y2=i Ak- By applying Proposition [521 the result follows. □ The structure of /3(y, Xjgp). For Y a closed subspace of Xjgp and T : Y — )• jCjgp we show that T = A/y.^^gp + 5* with 5:1"—)- Xjgp strictly singular. Proposition 5.5. Let y be a subspace of Xjgp and T : F — > Xjgp be a linear operator, such that there exists {a^fcjfceN a sequence in Y generating a Co spreading model and limsupdist(rxfc,Mxfc) > 0. Then T is unbounded. Proof. Passing, if necessary, to a subsequence, there exists 1 > 6 > 0, such that dist{Txk,M.Xk) > 5, for any /c S N. Since {xfcjfcgN generates a cq spreading model, it is weakly null. Set yk = Txk and assume that T is bounded. It follows that passing, if necessary, to a subsequence of {xfej^gN, then {yk}k(m also generates a cq spreading model. We may assume that {xA,.}A,.gN, as well as {yk}k&i are block sequences with rational coefficients. And lim^ = 1, as well as lim^ \\yk\\ = 1- If this is not the case pass, if necessary, to a further subsequence of {xfcjfcgN; such that both {xfcjfcgN and {yk}k&n are equivalent to some block sequences with rational coefficients {x'^jfcgN, {^fclfceN respectively, and more- over limfc ||x'^|| = 1, as well as lim^ ||?/^|| = 1. Set Y' = [{x'i^}ke^\ and T' : y — > Xjgp , such that T'x'^ — v'k- easy to check that T' is also bounded and dist(T'x'^, Rx'^^,) > 5' , for some 5' > 0. Set Ik = ran(ranxfc Uranyfc) and passing, if necessary, to a subsequence, we have that {/fcjfcgN is an increasing sequence of intervals of the naturals. We will choose {/fcjfcGN C W, such that fk{yk) > f , fk{xk) = and ran/fc C Ik, for all A; € N. The Hahn-Banach Theorem, yields that for all A; G N, there exists G -Bx*gp ! such that /^(y^) > S, fj^ixk) = and van f'^ C Ik-, for all € N. By the fact that jCjgp is reflexive, it follows that W is norm dense in -S;e*gp ) therefore there exists G W with \\fl — f'^W < j and ran/^' C Ik, for all A; G N. It follows that f'^iyk) > < 3 and /^'(x^) is rational, for ah ken. Furthermore, there exists gk G W, such that gk{xk) > 1 ~ 31 dki^k) is rational and ran gk C Ik, for all A; G N. Set fk = \{fk — gfc(xfcj fffc)- doing some easy calculations, it follows that the fk are the desired functionals. HEREDITARY INVARIANT SUBSPACE PROPERTY 33 By copying the proof of Proposition l4.10l for any fco S N, n € N and e > 0, there exists F a finite subset of the naturals with minF ^ fco and {ck}k&F such that (i) HkeF^^kXk is a (n,e) s.c.c. (ii) If z = 2" Y.k^F CkXk, then f ^ ^ 28. (iii) {xk}k£F is (4, {rifcjfcgjT') a-RIS and 2^" < riminF- (iv) There exists a functional / of type Iq with weight w{f) = n such that f{z) = 0, maxsupp/ > maxsuppz and if w = 2"^ J2keF '^kUk, then f {w)> ^. Arguing in the same way as in the proof of Theorem 15.41 for some n € N, we construct a sequence {zk}^^i and {gk}k=i such that {{zk, gk)}k=i 0" dependent and if Wk = Tzk, then gk{wk) > | and rang^ n ranwm = J2(, for k ^ m. Then / = ^ X]fc=i 9fc a functional of type II and ^||X]fc=i^fcil ^ Moreover, Proposition 15.31 yields that ^ |[ X]fc=i -Zfcll ^ It follows that ||T|| > Since n was randomly chosen, T cannot be bounded, a contra- diction which completes the proof. □ In [14], it is proven that if X is a hereditarily indecomposable complex Banach space, y is a subspace of X and T : y — ?■ X is a bounded linear operator, then there exists A G C, such that T — Xlyx '■ Y ^ X is strictly singular. Here we prove a similar result for Xjgp . Theorem 5.6. Let Y be an infinite dimensional closed subspace of Xjgp and T : Y ^ ^isp ^"6 a bounded linear operator. Then there exists A € M, such that T — XI^^^^^ : Y — )• Xjgp is strictly singular. Proof. If T is strictly singular, then evidently A = is the desired scalar. Otherwise, choose Z an infinite dimensional closed subspace of Y, such that T : Z ^isp ^^^o isomorphism. Choose {a^^l^gN a normalized sequence in Z generating a cq spreading model. Proposition 15.51 yields that liuik dist{T Xk,^Xk) = 0. Choose {Afcl^gN scalars, such that limfe ||Tj;fc — AfcXfcll = and A a limit point of {AfcjfcgN- We will prove that S = T — Xlyx is strictly singular. Towards a contradiction, suppose that this is not the case. Then there exists {y^lfceN a normalized sequence in Y generating a cq spreading model and 6 > 0, such that \\Syk\\ = ||(T - XI^^^^^ )yk\\ > 6, for all keN. As previously, we may assume that {xfcjjfcgp}, {yfelfeeN as well as {Syk}k€N are all normalized block sequences generating cq spreading models. By Proposition l5.5l and passing, if necessary, to a subsequence, there exists G M, such that lim^ \\Syk — nVkW = 0. Evidently / 0, otherwise we would 34 S.A. ARGYROS, P. MOTAKIS have that hnifc HS'yfcH = 0. Pass, if necessary, to a further subsequence of {yk}km, such that X]fcli \\Syk - mil < Observe that hm^ ||>S'xfc|| = and therefore we may pass, if necessary, to a subsequence of {xk}keN, such that X^fc^i 11*5*3;/; || < Arguing in the same manner as in the proof of Theorem 15.41 for some n G N construct {-Zfc}^" ^ ^^'^ {/fc}fc=i such that Z2k-~i is a hnear combination of {ykjken, Z2k is a hnear combination of {x^jfeeN and {{zk,fk)}lZi is a 1- dependent sequence. Set f = \ Z]fc=i /fc; which is a functional of type II in W. If Wk = Z2k-i - Z2k, Proposition [O] yields that ^\\ Ylk=i'"^k\\ ^ On the other hand, we have that -j^ rt 1 ^ -W^SWkW ^ -(II J^5z2fc-l|| - II X]^^2fc||) k=l k=l k=l ^ ~(llIZ/^^2fc-l|| - \\^{Sz2k-l -flZ2k-l)\\ - k=l k=l n 1/ 1/^*1 V^j^ 29|/i| 29|/x| > -{^Y.f^k-iiz2k-i] 232 ^ fc=i ^ _H _ IH > _H 2 4n ^ 4 It follows that 1 1 51 1 ^ where n was randomly chosen. This means that S is unbounded, a contradiction completing the proof. □ Strictly Singular Operators. In this subsection we study the action of strictly singular operators on Schauder basic sequences in subspaces of Xjgp . Proposition 5.7. Let Y be an infinite dimensional closed subspace of j£jgp and r : y — )• Xjgp be a linear bounded operator. Then the following asser- tions are equivalent. (i) T is not strictly singular. (ii) There exists a sequence {xfcjfceN in Y generating a cq spreading model, such that {Txk}kGN is not norm convergent to 0. Proof. Assume first that T is not strictly singular and let Z be an infinite dimensional closed subspace of Y, such that T\z is an isomorphism. Since any subspace of Xjgp contains a sequence generating a cq spreading model, then so does Z. Since T\z is an isomorphism, the second assertion is true. Assume now that there exists {xfcjfcgN a sequence in Y generating a cq spreading model, such that {Txk}keN does not norm converge to 0. By Proposition 15.51 and passing, if necessary to a subsequence, there exists A ^ 0, such that lim^ ||Txfc — Xxk\\ = 0. Passing, if necessary, to a further HEREDITARY INVARIANT SUBSPACE PROPERTY 35 subsequence, we have that X^^^i \\Txk — Axfc|| < oo. But this means that {xfcjfceN is equivalent to {Txk]k&i-, therefore T is not strictly singular. □ Proposition 5.8. Let Y be an infinite dimensional closed subspace of jCjgp and S, T : y — )• y be strictly singular operators. Let also be a weakly null sequence in Y generating an ii spreading model, such that {STxk}k&i does not norm converge to 0. Then passing, if necessary, to a subsequence, {STxk\k&n generates a cq spreading model. Proof. Towards a contradiction, assume that this is not the case. Then passing, if necessary to a subsequence, {xfcjfcgN, {Txk}^^-^;^ and {STxk\k&i all generate li spreading models. Set yk = Txk and Zk = STxk- We may assume that {a^^}, {yk}k&i as well as {zk\k&i are all seminormalized block sequences. Set Ik = ran(ranxfc U ranyjfc U ranz^) for all fc G N and pass, if necessary, to a subsequence of {xfc}fc6N, such that {/fcjfcgN is an increasing sequence of intervals of the naturals and set jk = minljt. It is not hard to see that if Xlfce-F ^^^h i^ ^ C'^' ^) basic s.c.c. then J2keF (^kXk,J2k€F (^kVk as weh as J2keF'^kZk, are ah (n,2e) s.c.c. Moreover either a[{xk}k) > or /3[{xk}k) > and the same is true for {Vkjkm and {^felfceN- Arguing in the same way as in the proof of Proposition 13.51 there exists c > 0, such that for any n, A^q G N and e > 0, there exists F a finite subset of the naturals with minF ^ and {ck}keFi such that (i) X] X] '^'^y^' ^""^ X] '^^^'^ ^"^^ ('^'^^ fceF fcgF keF (ii) II ^'=^'^11 > ^' II ^'=^'=11 > ^' ^^'^ II '^'=^'=11 > ^ keF keF keF Choose {nfcjfcgN strictly increasing naturals, {-FfcjfcgN increasing subsets of the naturals and {ci}ii^Fk, such that if x'f^ = 2"*= X^igF^ (^i^hTx'^ = y'k = ^^'^ EigFfc CiVi and STx'^ = 4 = 2"*= EieF^ ^i^i, then {x'^jfceN, {yU^eN and {z^jfcgN are all seminormalized and satisfy the assumptions of Proposition [3S1 Therefore a({x'^}fc) = 0,a[{y'i^}k) = and a[{z'f^}k) = 0. We will show that /3{{x'f^}k) > 0, as well as (3{{y'^}k) > 0. If P{{Xk}k) = 0, then Proposition 14.81 yields that passing, if necessary, to a subsequence, {x'j^}km generates a cq spreading model and since {y^j^gN is seminormalized and Tx'i^ = y'^, so does {y^jfcgN- Proposition 15.71 yields a contradiction. For the same reasons we also conclude that /3({yfc}fc) > 0. Using the same arguments, we may construct {t^fcjfcgN a block sequence of {x'^jfcgN, such that both {iffcjfcgN and {Twk}k&i are seminormalized and satisfy the assumptions of Corollary 13.171 Therefore we conclude that a{{wk}k) = ^,P{{wk}k) = Q,a{{Twk}k) = and li{{Twk}k) = 0. Propo- sitions 14.81 and 15.71 yield a contradiction and this completes the proof. □ 36 S.A. ARGYROS, P. MOTAKIS The Invariant Subspace Property. Theorem 5.9. Let Y be an infinite dimensional closed subspace of j£jgp and Q, S,T : Y ^ Y he strictly singular operators. Then QST is compact. Proof. Since Xjgp is reflexive, it is enough to show that for any weakly null sequence {xfcjfcgN) we have that {QSTxk}keN norm converges to zero. Pass, if necessary, to a subsequence again denoted by {xfcjfceN, that generates some spreading model, which is, as we have shown, either equivalent to ii, or to cq. Assume first {xk}ken generates a cq spreading model. If {Txk}keN is not norm convergent, then it has a subsequence generating a cq spreading model as well. Proposition 15.71 vields a contradiction. If {xfclfcgN generates an ii spreading model and {STxk}k€N is not norm convergent, then Proposition 15. 81 vields that passing, if necessary, to a subse- quence of {xfcjfcgN, {STxk}keN will generate a cq spreading model. Arguing as in the previous case, we conclude that {QSTxk}keN is norm convergent and this completes the proof. □ Corollary 5.10. Let Y be an infinite dimensional closed subspace of J£jgp and S : y — 7- y be a non zero strictly singular operator. Then S admits a non-trivial closed hyperinvariant subspace. Proof. Assume first that S''^ = 0. Then it is straightforward to check that ker 5 is a non-trivial closed hyperinvariant subspace of S. Otherwise, if 5^ ^ 0, then Theorem 15.91 yields that is compact and non zero. Since S commutes with its cube, by Theorem 2.1 from [27], it is enough to check that for any a, /3 € M such that /? 7^ 0, we have that {al — 5)^ + 7^ 0. Since S is strictly singular, it is easy to see that this condition is satisfied. □ Corollary 5.11. Let Y be an infinite dimensional closed subspace of jCjgp and T : y — )• y be a non scalar operator. Then T admits a non-trivial closed hyperinvariant subspace. Proof. Theorem 15.61 yields that there exist A G M, such that S = T — XI is strictly singular, and since T is not a scalar operator, we evidently have that S is not zero. By Corollarv 15.101 it follows that 5 admits a non-trivial closed hyperin- variant subspace Z. It is straightforward to check that Z also is a hyperin- variant subspace for T. □ In the final result, which is related to Proposition 3.1 from [5], we show that the "scalar plus compact" property fails in every subspace of Xjgp . Proposition 5.12. Let Y be an infinite dimensional closed subspace of Xjgp . Then there exists a strictly singular, non compact operator S : Y Y. In fact, if S(Y) is the space of strictly singular operators on Y, then S(Y) is non-separable. HEREDITARY INVARIANT SUBSPACE PROPERTY 37 Proof. By Corollary I4.1H there exists a sequence {xfcjfcgN in ^ that gener- ates a spreading model equivalent to cq, say with an upper constant ci and by Corollary 14. 131 there exists a sequence {x1}km in ^* that also generates a spreading model equivalent to cq, say with an upper constant C2. There- fore {xfcjfcgN and {x^jfcgN are weakly null and we may assume that they are Schauder basic and that dim(y/[xfe]fc) = oo. We may also assume that there exist {zk}k€N in Y such that {x^j/cgN is almost biorthogonal to {zk}keN- For e > 0, set = Choose a strictly increasing sequence of naturals {gjjjgN, such that qj ^ M]^/2^+i- Set S : Y ^ Y, such that = I]fcii^gfe(^)^fc- Then: (i) S is bounded and non compact. (ii) S is strictly singular. We first prove that it is bounded. Let x €Y, \\x\\ = 1, x* G Y*, \\x*\\ = 1. For j ^ 0, set Bj = {k £ N : 1/2^+^ < \x*{xk)\ < 1/2^}. Since {xfejfcgN generates cq as a spreading model, it follows that Bj ^ Mi/2j+i ^ qj- Set Cj = {k G Bj : k = \ Cj- Evidently #-Dj ^ j and it is easy to see that #{qk '■ k G Cj} ^ mm{qi^ : k € Cj}, therefore, since {x^j^gN generates a spreading model equivalent to cq, it follows that I ^ x*{xk)x*g^{x)\ < C2max{|3;*(xfc)| : k G Cj} Therefore | EfceB, ^ C2 max{|2;*(xfc)| : k e Cj}+j/2^ ^ C2/2-' -\rj/2^. From this it follows that oo • _j_ IIS'xIl ^ > : ||x|| II II ^ 23 " " j=0 The fact that S is non compact follows easily if you consider the almost biorthogonals {zk}keN of {a^g^,}fceN- Then {^fcjfcgN is a seminormalized se- quence in Y and {SzfclfcgN does not have a norm convergent subsequence. We now prove that S is strictly singular. Suppose that it is not, then there exists A / such that T = S — XI is strictly singular. Since XI is a Fredholm operator and T is strictly singular, it follows that S = T + XI is also a Fredholm operator, therefore dim(y/S'[y]) < oo. The fact that S[Y] C [xk]k and dim(Y/[xk]k) = oo yields a contradiction. Moreover, for any further subsequence {xl}keL of {x*^}keN^ if we set Slx = J2k&L^ki^)^k, then 5*^ satisfies the same conditions. This yields that S{Y) contains an uncountable e-separated set and is therefore non- separable. □ The last proof actually yields that if Y is an infinite dimensional closed subspace of Xjgp , then the space of strictly singular, non-compact operators of Y is non-separable. 38 s.a. argyros, p. motakis References [1] D. E. Alspach and S. A. Argyros, Complexity of weakly null sequences, Dissertationes Math. (Rozprawy Mat.) 321 (1992), 44. [2] S. A. Argyros, K. Beanland and P. Motakis, The strictly singular operators in Tsirelson like reflexive spaces (in preparation). [3] S. A. Argyros and I. Deliyanni, Banach Spaces Of The Type Of Tsirelson ,arXiv:math/920 7206 (1992). [4] S. A. Argyros and I. Deliyanni, Examples of asymptotic l\ Banach spaces, Trans. Amer. Math. Soc. 349, no. 3 (1997), 973-995. [5] S. A. Argyros, I. Deliyanni and A. G. Tolias, Strictly singular non-compact diagonal operators on HI spaces, Houston J. Math. 36, no. 2 (2010), 513-566. [6] S. A. Argyros, G. Godefroy and H. P. Rosenthal, Descriptive set theory and Ba- nach spaces. Handbook of the Geometry of Banach Spaces, Vol. 2, North-Holland, Amsterdam (2003), pp. 1007-1069. [7] S. A. Argyros and R. G. Haydon, A hereditarily indecomposable Coo-space that solves the scalar-plus-compact problem. Acta Math. 206, no. 1 (2011), 1-54. [8] S. A. Argyros and A. Tolias, Methods in the theory of hereditarily indecomposable Banach spaces. Memoirs of the American Mathematical Society 170 (2004), vi-|-114. [9] M. Aronszajn and K. T. Smith, Invariant subspaces of completely continuous opera- tors, Ann. of Math. 60 (1954), 345-350. [10] J. Bourgain and F. Delbaen, A class of special Loo spaces. Acta Math. 145, no. 3-4 (1980), 155-176. [11] P. G. Casazza, W. B. Johnson and L. Tzafriri, On Tsirelson's space, Israel J. Math. 47, no. 2-3 (1984), 81-98. [12] P. G. Casazza and T. Shura, Tsirelson's Space, Spriger Lecture Notes 1363 (1989). [13] P. Enflo, On the invariant subspace problem for Banach spaces. Acta Math. 158, no. 3-4 (1987), 213-313. [14] V. Ferenczi, Operators on subspaces of hereditarily indecomposable Banach spaces. Bull. London Math. Soc. 29, no. 3 (1997), 338-344. [15] T. Figiel and W. B. Johnson, A uniformly convex Banach space which contains no Ip, Compositio Math. 29 (1974), 179-190. [16] W. T. Gowers and B. Maurey, The unconditional basic sequence problem, J. Amer. Math. Soc. 6, no. 4 (1993), 851-874. [17] R. C. James, Bases and refelexivity of Banach spaces, Ann. of Math. (2) 52 (1950), 518-527. [18] V. I. Lomonosov, Invariant subspaces of the family of operators that commute with a completely continuous operator, (Russian) Funkcional. Anal, i Prilozen. 7, no. 3 (1973), 55-56. [19] V. D. Milman, Operators of class Co and Cq, Teor. Funkcii Funkcional. Anal, i Prilozen. No 10 (1970), 15-26. [20] E. Odell and Th. Schlumprecht, On the richness of the set ofp's in Krivine's theorem. Operator Theory, Advances and Applications 77 (1995), 177-198. [21] E. Odell and Th. Schlumprecht, A Banach space block finitely universal for monotone bases, Trans. Amer. Math. Soc. 352, no. 4 (2000), 1859-1888. [22] F. P. Ramsey, On a problem of formal logic, Proc. London Math. Soc. Series 2, 30 (1930), 264-286. [23] C. J. Read, A solution to the invariant subspace problem, Bull. London Math. Soc. 16, no. 4 (1984), 337-401. [24] C. J. Read, A solution to the invariant subspace problem on the space l\. Bull. London Math. Soc. 17, no. 4 (1985), 305-317. [25] C. J. Read, Strictly singular operators and the invariant subspace problem, Studia Math. 132, no. 3 (1999), 203-226. HEREDITARY INVARIANT SUBSPACE PROPERTY 39 [26] Th. Schlumprecht, An arbitrarily distortable Banach space, Israel J. Math. 76, no. 1-2 (1991), 81-95. [27] G. Sirotkin, A version of the Lomonosov invariant subspace theorem for real Banach spaces Indiana Univ. Math. J. 54, no. 1 (2005), 257-262. [28] B. S. Tsirelson, Not every Banach space contains £p or co. Functional Anal. Appl. 8 (1974), 138-141. National Technical University of Athens, Faculty of Applied Sciences, De- partment OF Mathematics, Zografou Campus, 157 80, Athens, Greece E-mail address: sargyros@math.ntua.gr, pmotakis@central.rLtua.gr